Health check and auto-restart all Pollinations GPU services (Flux on Vast.ai, Z-Image on RunPod, LTX-2 on GH200, Klein on RunPod, legacy image on OVH, Sana on Oracle Cloud). Use with /loop for recurring checks.
SkillScore89Maintained
Automatic
◆DevOps & SRE pack
Last verified 10/8/2026
What it changes
Passed 17 of 19 test cases with the skill, and 9 of 19 without it.
Without skill47%
With skill89%
Checks it passed that the model alone missed
Queries model_health.pipe for failure and latency metrics
Queries recent_server_errors.pipe for detailed failure logs
Notes that KLEIN_URL is used for environments lacking the KLEIN_VPC binding
and 8 more in the full report
8 fixed9 passed either way2 still failing0 broke3 errored
Pass-rate lift
+42 pp
Tokens
-16%
Agent turns
0%
Response time
-37%
Model gemini-3.6-flashMeasured 10/8/2026v417 of 22 cases passed with skill · 19 comparable
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 8 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 8
Prompt
We need to fetch current active instance registrations from https://gen.pollinations.ai/register. What HTTP authentication header or token is required?
What was checked, without and with the skill
Checked
Without skill
With skill
States that listing registrations from https://gen.pollinations.ai/register requires no token
Missed
Passed
case-16 · gemini-3.6-flash · 10/8/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md21 linessingle file
Monitor GPU services
Read the repository AGENTS.md, the current GPU inventory, and the affected model's code before acting. Discover live instances and routes from the provider and Gen; documented instance IDs, addresses, prices, and retired services are historical until verified. Keep scheduling and alert thresholds in the task that invokes this skill.
Check the current fleet
Reconcile the inventory with current Gen routes, live provider instances, and GET https://gen.pollinations.ai/register (listing needs no token). Check every active GPU route; flag a routed service missing from the inventory, and skip historical services without an active route.
Check the model health endpoint for final failures, primary failures, fallback rescues, and latency. Use recent server errors and Worker logs for a specific failure. A successful fallback can hide an unhealthy GPU.
For a suspected worker, compare its direct health and authenticated generation with a real request through gen.pollinations.ai. Check the GPU process, queue, tunnel, and logs on the verified host. A registry heartbeat does not prove the data path works; a failed probe credential does not prove the backend is down. Some tunnel hostnames return a bot-protection 403 to Python-urllib/*; retry health probes with curl or another user agent before declaring failure.
Check how the workload is routed before interpreting a missing registry entry. Production Klein uses the KLEIN_VPC binding to a private Vast tunnel, so it does not appear in the heartbeat registry. KLEIN_URL is for environments without that binding.
Use an existing authorized test credential without printing it. Keep probes bounded, and compare failures with recent demand and provider fallback before declaring an outage.
Recover and verify
If the monitoring task authorizes repair of an existing worker, identify the exact live instance and use its checked-in setup or restart procedure. Recheck the process, direct endpoint, public Gen request, and final error rate after the restart.
For new capacity, a replacement instance, routing or tunnel changes, use the Vast fleet skill. Its prepare and promote steps have separate scope and approval requirements.
For any credential deployment or secret change, follow the separate approval gate in AGENTS.md. Production Cloudflare Worker deployments and secret synchronization run through Deploy / Cloudflare production from the production branch; never push production Worker secrets from a local session.
If the current host or approved recovery path cannot be verified, stop before changing production and report the impact and required next action.
Report each active workload's route, public and backend health, fallback use, action taken, and post-action verification. Mark retired or unverified services explicitly rather than running historical restart commands.